How Listenr got to 27 GitHub stars, one burst of motivation at a time

Developer Rebreda's open-source Listenr project reached 27 GitHub stars as of October 2026, up from 1 star at its first commit in October 2025, according to a first-person account of the project's growth. Listenr grew out of contributions to Lemonade, AMD's open-source local AI server, and uses Whisper plus custom voice activity detection and streaming to collect and improve speech-to-text transcripts on consumer hardware. The author credits community participation in the Lemonade Discord, a Firefox sidebar-chat integration post, intermittent commits, and an unexpected FOSS challenge win for the project's traction.

Development https://quickthoughts.ca/categories/development/ Opinion https://quickthoughts.ca/categories/opinion/ 6 min read How Listenr got to 27 GitHub stars, one burst of motivation at a time Listenr is the most popular open source project I have started. This is a little background to how it got to 27 stars thanks to Lemonade, a Firefox post, a lot of off-and-on commits, and a FOSS challenge win I did not see coming. Listenr https://github.com/Rebreda/listenr has 27 stars as of today. Admittedly that is not a big number compared to massively popular AI projects out there now Pi, hermes, etc but it dwarfs all my other open source projects, by a lot. I figure there are a lot of other dev's out there that are in the same boat as me that might find it interesting to see how I managed to get a project with a little bit of traction/usage. Moreover, it's been a fun ride, and I wanted to write down how it actually happened while I still have the data in front of me. Stars since first commit: 2025-10 1 2026-01 2 2026-03 6 2026-05 8 2026-06 15 2026-07 21 2026-08 22 2026-09 26 2026-10 27 First things first I mean, it's obvious - give your repo a star - it's free and 1 is infintely greater than 0. Plus if you don't like your own repo, why would others? Motivation is key but fickle Ideas/projects that others use generally need to solve some common problem. Listenr came out of contributing to Lemonade https://github.com/lemonade-sdk/lemonade , the open source local AI server from AMD. I'm huge proponent of running models that are feasible on consumer hardware and wanted a way to better utilize my AMD hardware. Whats more is the team behind lemonade was/is really open and willing to support contributors and actually ship PRs - which is not always the case for larger company-driven open source projects. So right off the bat, I was able to harness some motivation by participating with a community. I also like Firefox I swear this all ties back to the stars eventually and its sidebar chat, even though it is a conceptually simple integration. Firefox lets you point that sidebar at a local model, but not easily. So I wrote a post on making it work with Lemonade https://quickthoughts.ca/posts/firefox-chatback-lemonade-sdk and sent a few PRs to make the two fit together better. The project needed some web love, and that happened to be an area I actually know: web development, audio APIs and an afinity for STT. It was really satisfying to see it all just work together, and it got me talking with the team on the Lemonade Discord. In the background, I had always wanted to actually fine-tune ASR/voice models to better work for how I speak and ideally be more efficent. I had built out a few projects to do collect my own voie data and transcribe them manually in the past but it was just too much time to do it. However, by 2025, AI models got small enough to run locally, they could do the grunt work of improving transcripts reliably enough to get over that hump which let me to create the v1 of listenr. The goal was to use whisper + custom VAD/streaming to kick off the transcription collection/improvement process. The end result was predictable - the Voice Activity Detection VAD and the ergnomics of real-time transcription were still painful even with whisper and small models. Whisper hallucinates a lot when theres silence or background noise, meaning transcriptions triggered by my custom VAD logic where often filled with random text and needed more processing kinda defeating the whole point . Long story short, it eventually dawned on me that lemonade also provided their own OAI competiable endpoints that handled VAD and streaming already. I immediately got to work and simply said, their implementation was much better than mine. Still I eventually lost motivation and the project kind of went into hibernation and no one used it - something that I find is typical with a lot of OS projects out there. The challenge Things didn't pick up again until I learned of a lemonade developer challenge https://www.amd.com/en/developer/resources/technical-articles/2026/join-the-lemonade-developer-challenge.html still on-going as of Oct 2026 that AMD was putting together in which winners would get a sick HP Zbook if their project was chosen. That immediately re-piqued my interest and gave me the carrot to take listenr from "neat but niche" to a full end-to-end local-first fine-tuning pipeline that could easily run on AMD hardware. This ended up being a pretty common issue others ran into - not being able to full take advantage of rocm/AMD gpus for fine-tuning because of hard-to-setup drivers/packages and also not being able to get enough data to fine-tune on. I eventually submitted the project to the challenge, and wrote about it in my blog. Along the way, I collected a lot more voice data and really got into fine-tuning, learning a lot along the way. It was rewarding to actually build out something novel, by building on existing FOSS projects instead of reinventing everything myself . After a few months, I learned listenr actually won Which was completely unexpected and also very cool. The laptop has been an absolute unit and massive upgrade when it comes to running medium-sized models and intensive training sessions. Since then, the project has continue to pick up stars thanks in part of my blog and referral traffic from other posts. The few pivots of listenr Listenr has had a few lives. Roughly: 1. October 2025: get something that kind of works. A command line listener where Silero VAD cuts the mic into clips, Whisper transcribes them locally, and an LLM in the loop cleans up small transcription mistakes. 2. February 2026: rebuild it on Lemonade, and turn it into a data collector. Self-hosted Whisper and Ollama were swapped for Lemonade Server, the web UI was deleted, and build dataset.py showed up. Swapped out my own custom VAD code. 3. March 2026: fine-tuning. LoRA fine-tuning on AMD with ROCm and Docker, the biggest single change in the repo at about 4300 lines. 4. May 2026: Listenr wins the lemonade challenge https://github.com/lemonade-sdk/lemonade/issues/2035 5. July 2026: Added a way to leverage public datasets via Mozilla Data Collective and hugging face. 6. August 2026: Public to PyPI, add Moonshine models alongside Whisper, and then ton of actual e2e training using the project. That was also the week fine-tuning made it worse https://quickthoughts.ca/posts/finetuning-made-it-worse . That is about a hundred commits and 25 PRs, with plenty of failures in between. Motivation comes in waves Development has been very off-and-on. The commits by month make that obvious: 2025-10 7 2026-02 14 2026-03 22 2026-04 1 2026-06 18 2026-07 2 2026-08 42 Real life doesn't make it easy to spend hours getting into the weeds on fine-tuning or fighting annoying dependency version issues. What's next? I couldn't really tell you - I think I'll keep plugging away at making the data collection side more useful via things like dictatr and keeping up with the latest in ASR models. I think I'll hopefully be able to get another wave of motivation by more people sharing how they use the project. I would love to know what people want from Listenr, and how their data collection and training runs are going. If you have tried it, open a discussion https://github.com/Rebreda/listenr/discussions or issue , even if it did not work. The bottleneck is still natural speech data. Recording it on purpose is awkward, and that is what pushed me to start dictatr https://github.com/Rebreda/dictatr : dictation I would use anyway, that collects the data as a side effect.